Added disable-handle-pseudo-memory-ops option.

ispc.cpp

@@ -284,6 +284,7 @@ Opt::Opt() {
     fastMaskedVload = false;
     unrollLoops = true;
     disableAsserts = false;
+    disableHandlePseudoMemoryOps = false;
     disableBlendedMaskedStores = false;
     disableCoherentControlFlow = false;
     disableUniformControlFlow = false;

ispc.h

@@ -227,6 +227,12 @@ struct Opt {
     /** Indicates whether assert() statements should be ignored (for
         performance in the generated code). */
     bool disableAsserts;
+    
+    /** If enabled, the various __pseudo* memory ops (gather/scatter,
+        masked load/store) are left in their __pseudo* form, for better
+        understanding of the structure of generated code when reading
+        it. */
+    bool disableHandlePseudoMemoryOps;
 
     /** On targets that don't have a masked store instruction but do have a
         blending instruction, by default, we simulate masked stores by

main.cpp

@@ -88,6 +88,7 @@ static void usage(int ret) {
     printf("        fast-masked-vload\t\tFaster masked vector loads on SSE (may go past end of array)\n");
     printf("        fast-math\t\t\tPerform non-IEEE-compliant optimizations of numeric expressions\n");
 #if 0
+    printf("        disable-handle-pseudo-memory-ops\n");
     printf("        disable-blended-masked-stores\t\tScalarize masked stores on SSE (vs. using vblendps)\n");
     printf("        disable-coherent-control-flow\t\tDisable coherent control flow optimizations\n");
     printf("        disable-uniform-control-flow\t\tDisable uniform control flow optimizations\n");
@@ -254,6 +255,8 @@ int main(int Argc, char *Argv[]) {
 
             // These are only used for performance tests of specific
             // optimizations
+            else if (!strcmp(opt, "disable-handle-pseudo-memory-ops"))
+                g->opt.disableHandlePseudoMemoryOps = true;
             else if (!strcmp(opt, "disable-blended-masked-stores"))
                 g->opt.disableBlendedMaskedStores = true;
             else if (!strcmp(opt, "disable-coherent-control-flow"))

opt.cpp

@@ -201,8 +201,10 @@ Optimize(llvm::Module *module, int optLevel) {
         // them into something that can actually execute.
         optPM.add(llvm::createPromoteMemoryToRegisterPass());
         optPM.add(CreateGatherScatterFlattenPass());
-        optPM.add(CreateLowerGatherScatterPass());
-        optPM.add(CreateLowerMaskedStorePass());
+        if (g->opt.disableHandlePseudoMemoryOps == false) {
+            optPM.add(CreateLowerGatherScatterPass());
+            optPM.add(CreateLowerMaskedStorePass());
+        }
         optPM.add(CreateIsCompileTimeConstantPass(true));
         optPM.add(llvm::createFunctionInliningPass());
         optPM.add(CreateMakeInternalFuncsStaticPass());
@@ -282,8 +284,10 @@ Optimize(llvm::Module *module, int optLevel) {
         optPM.add(CreateLowerMaskedStorePass());
         if (!g->opt.disableGatherScatterOptimizations)
             optPM.add(CreateGatherScatterImprovementsPass());
-        optPM.add(CreateLowerMaskedStorePass());
-        optPM.add(CreateLowerGatherScatterPass());
+        if (g->opt.disableHandlePseudoMemoryOps == false) {
+            optPM.add(CreateLowerMaskedStorePass());
+            optPM.add(CreateLowerGatherScatterPass());
+        }
         optPM.add(llvm::createFunctionInliningPass());
         optPM.add(llvm::createConstantPropagationPass());
         optPM.add(CreateIntrinsicsOptPass());